Multimodal Dataset for Detecting Audience–Influencer Parasocial Relationships
Description
The dataset contains 20,000 audience–content observations associated with 120 YouTube influencers across six content categories: beauty and fashion, fitness and health, travel and lifestyle, food, technology and gaming, and entertainment and education. The sample is balanced across micro-, mid-tier, macro-, and mega-influencers. Each observation combines available textual, visual, behavioral, and temporal information. These data may include audience comments, influencer captions and replies, transcripts, sampled visual content, engagement indicators, interaction frequency, continuity of participation, influencer responsiveness, and longitudinal interaction patterns. A stratified subset of 5,000 observations was manually annotated using a theory-grounded coding framework. The annotations capture overall parasocial relationship strength and seven dimensions: perceived intimacy; identification; emotional attachment; trust; loyalty; personal knowledge; and perceived reciprocity. Each dimension was evaluated using four ordinal categories: absent, weak, moderate, and strong. The remaining 15,000 observations constitute an unlabeled application sample intended for applying and evaluating finalized computational models. The dataset also includes aggregated validation information from 800 survey respondents associated with 40 influencers, with 20 respondents evaluating each influencer. The survey component provides an independent criterion for examining convergence between computationally inferred and self-reported parasocial relationship strength. The dataset can be used for research on multimodal machine learning, natural language processing, computational social science, influencer marketing, explainable artificial intelligence, ordinal classification, construct validation, and the computational measurement of latent social and psychological constructs. Direct personal identifiers were removed before analysis, and influencer, post, and audience identifiers were replaced with study-specific pseudonymous codes. The data are intended exclusively for academic and non-commercial research conducted in accordance with applicable ethical and data-protection requirements.
Files
Institutions
- Technical University of LiberecLiberecký kraj, Liberec